Artificial intelligence algorithms in assisted reproduction: differences in the evaluation of embryos from fresh or vitrified donor oocytes

Abstract Study question Do the Embryoscope and EmbryoscopePlus systems' built-in software evaluate differently embryos from vitrified or fresh oocytes? Summary answer The embryo scores provided by artificial intelligence (AI) algorithms are lower for embryos originated from vitrified/warmed oocytes than for those that come from fresh oocytes. What is known already Time-lapse technology has enabled embryologists to develop several AI-based tools that allow the automation and standardization of embryo quality assessment. EmbryoScope and EmbryoScopePlus systems include deep learning-based models using only time-lapse image sequences (iDAScore v1.0 and v.2.0) and the machine learning-based model (KIDScore D5 v3) which classifies embryos in categories based on cleavage time points and blastocyst appearance. It has been reported that fertilization, pregnancy and implantation rates in vitrified oocytes are comparable to those of fresh oocytes. To our knowledge, this is the first study to test whether artificial intelligence tools assess embryos derived from these oocytes differently. Study design, size, duration This retrospective study includes the analysis of 34950 videos of embryos from fresh and vitrified donor oocytes. Embryos were cultured on EmbryoScope and EmbryoScopePlus Time Lapse systems and morphological and morphokinetic parameters were automatically detected by EmbryoViewer software. Senior embryologists routinely evaluated the embryos according to the ASEBIR criteria. In addition, embryos were graded using IDAScore v1, IDAScore v2 and KIDScoreD5 algorithms in different scores from low to high quality (1-9.9). Participants/materials, setting, methods The scores of embryos from fresh or vitrified donor oocytes obtained with the different algorithms were compared. Specifically, 21533 embryos from fresh and 12710 from vitrified oocytes with the IDAScore v1; 21335 from fresh and 12705 from vitrified with the IDAScore v2; and 13544 from fresh and 6559 from vitrified with the KIDScore v3. In addition, the ASEBIR evaluation performed by embryologists was also compared between embryos from fresh (21968) and vitrified oocytes (12980). Main results and the role of chance Mean scores of embryos coming from fresh oocytes from IDAScore v1, IDAScore v2 and KIDScore v3 were 5.34 ± 2.91, 3.48 ± 2.93, 5.13 ± 2.1 respectively. And mean scores of those from vitrified oocytes were 5.09 ± 2.83, 3.1 ± 2.71, 4.99 ± 2.01. Thus, mean embryo scores obtained by the 3 artificial intelligence algorithms were significantly higher* for embryos from fresh oocytes compared to those from vitrified oocytes. However, this lower embryo assessment of those from vitrified oocytes was not detected when comparing embryo quality according to ASEBIR classification. We hypothesized that it might be because several embryos that had not reached blastocyst stage on D6 were included in this comparison. Comparisons were repeated excluding these embryos, mean scores of embryos from fresh oocytes from IDAScore v1, IDAScore v2 and KIDScore v3 were 7.31 ± 2.12, 5.15 ± 2.79, 5.18 ± 2.07 respectively. Mean scores of those from vitrified oocytes were 7.14 ± 2.12, 4.75 ± 2.70, 5.01 ± 2.01. Again, mean scores of embryos from fresh oocytes were significantly higher* than those from vitrified oocytes. In this analysis, this significant difference* in embryo score was also detected according to the ASEBIR classification. *p<.05 Limitations, reasons for caution This project is limited by its retrospective and single-center nature. A multicenter study could be performed to determine whether these minimal but significant differences in the evaluation of embryos from fresh or warmed oocytes have any impact on clinical outcome. Wider implications of the findings This study shows that embryos from vitrified oocytes have significantly lower scores than those from fresh oocytes. The embryo assessment performed in this study is based on AI models that allow their instantaneous, non-invasive, accurate and objective evaluation. Thus, the impact of these differences on clinical outcome, should be analyzed. Trial registration number not applicable